Multiscale edge detection based on Gaussian smoothing and edge tracking

The human vision is usually considered a multiscale, hierarchical knowledge extraction system. Inspired by this fact, multiscale techniques for computer vision perform a sequential analysis, driven by different interpretations of the concept of scale. In the case of edge detection, the scale usually relates to the size of the region where the intensity changes are measured or to the size of the regularization filter applied before edge extraction. Multiscale edge detection methods constitute an effort to combine the spatial accuracy of fine-scale methods with the ability to deal with spurious responses inherent to coarse-scale methods. In this work we introduce a multiscale method for edge detection based on increasing Gaussian smoothing, the Sobel operators and coarse-to-fine edge tracking. We include visual examples and quantitative evaluations illustrating the benefits of our proposal.